CartPilot AI

A Gemini-powered AI shopping agent that turns natural-language requests into relevant product recommendations and enables seamless checkout through PayPal Sandbox.

Comment

Inspiration

Online shopping can become time-consuming when users have to search with specific keywords, manually apply filters, and compare multiple products to find what they need.

I wanted to make product discovery more natural. Instead of requiring users to know exactly what to search for, what if they could simply describe what they wanted in everyday language?

That idea inspired CartPilot AI: an AI-powered personal shopping agent that connects natural-language understanding, relevant product recommendations, and a payment checkout experience in one application.

What it does

CartPilot AI transforms natural-language shopping requests into structured product requirements and relevant recommendations.

Users can:

  • Describe their shopping needs using everyday language.
  • Explore suggested prompts to discover products.
  • Receive AI-assisted product recommendations based on their requirements.
  • Browse a catalog of 25 demo products and explore product details.
  • Add products to their cart and review their selections.
  • Proceed through a PayPal Sandbox checkout flow.

Google Gemini helps interpret shopping requests, while the backend applies structured filtering and relevance scoring to match products against the catalog.

How we built it

CartPilot AI uses a full-stack architecture:

  • Frontend: React, TypeScript, and Vite.
  • Backend: Python and FastAPI.
  • AI integration: Google Gemini API for natural-language shopping intent understanding.
  • Product discovery: A structured 25-product demo catalog with deterministic filtering and relevance ranking.
  • Payments: PayPal Sandbox for order creation, approval, verification, and payment capture.

The frontend communicates with the FastAPI backend through API endpoints. When a user submits a shopping request, the backend uses Gemini to interpret the intent and then applies product-matching logic to identify relevant catalog items.

For checkout, the backend creates a PayPal Sandbox order. The user approves the order through PayPal, returns to CartPilot AI, and the backend verifies and captures the payment.

The application is deployed publicly, with the frontend hosted on Vercel and the backend hosted on Render.

Challenges we ran into

One of the main challenges was converting flexible natural-language requests into reliable product matches.

Users can describe the same requirement in different ways, use incomplete phrases, or combine multiple constraints. We worked on intent matching, product-type preservation, filtering, and relevance-ranking regressions to make the recommendations more consistent.

Another challenge was integrating the PayPal checkout lifecycle correctly. Creating an order is not the same as completing a payment. We had to handle the approval redirect, verify the approved order, and capture the payment successfully.

We also worked through deployment configuration, environment variables, API communication, and frontend-backend integration to make the application function outside the local development environment.

Accomplishments that we're proud of

  • Built and deployed a complete full-stack AI shopping application.
  • Integrated Google Gemini for natural-language shopping intent understanding.
  • Implemented structured product filtering and relevance ranking.
  • Connected PayPal Sandbox to a complete order approval and payment capture flow.
  • Successfully tested the end-to-end Sandbox checkout, including transaction verification.
  • Developed a backend test suite with 97 passing tests in the verified test run.
  • Published the source code with an MIT License.

Most importantly, CartPilot AI demonstrates a complete shopping journey—from describing a need in natural language to receiving recommendations and completing a Sandbox checkout.

What we learned

Through this project, we learned that building an AI-powered application requires more than connecting a language model to a user interface.

Reliable AI products need structured backend logic, predictable data handling, appropriate error handling, and clear integration boundaries.

We also gained practical experience with Gemini API integration, FastAPI development, product relevance ranking, PayPal order management, payment capture, deployment, and end-to-end testing.

What's next for CartPilot AI

We envision evolving CartPilot AI into a more intelligent and personalized shopping assistant.

Potential improvements include:

  • Expanding the product catalog with real product data.
  • Improving recommendation quality and personalization.
  • Supporting product comparisons across multiple attributes.
  • Adding user accounts, shopping history, and saved preferences.
  • Exploring additional payment and commerce integrations.
  • Improving recommendation explanations and shopping assistance.

Our goal is to make online shopping more conversational, intelligent, and convenient.

Project: CartPilot AI
Developer: Plabon Basak
Payment environment: PayPal Sandbox — no real money is charged or transferred.

Built With

Share this project:

Updates

Submission history